Deep learning-based small bubble detection in research reactor water pool using consecutive multi-frame

Abstract

The safety and reliability of nuclear reactors are crucial, necessitating constant and precise monitoring. Identifying small bubbles in the reactor's water pool is a crucial component of this monitoring, as their presence can signal potential issues inside the reactor. These bubbles are challenging to detect with the naked eye due to their small size and similarity to the background. Furthermore, single-frame deep learning models find it challenging to detect bubbles because of limited pixel representation of small objects, absence of temporal context, and variations in bubble appearance, resulting in lower accuracy. In this paper, we propose a method that utilizes consecutive multi-frame analysis to enhance the accuracy of bubble detection. The synthetic bubble datasets were generated using high-resolution CCTV video from the HANARO, and a multi-frame based bubble detectors were developed to detect bubbles. The multi-frame bubble detector exhibited a 21.33% higher AP on synthetic bubbles compared to the single-frame bubble detector, along with a 1.5% improvement in AP on real bubbles. The multi- frame approach demonstrated an improvement in detection accuracy compared to single-frame analysis by reducing false positives and more effectively capturing the dynamic behavior of bubbles. This research advances safety monitoring in nuclear reactors and highlights the effectiveness of deep learning in handling complex surveillance scenarios.